Abstract
Audio Source Separation is an interesting problem, which gives us the power to separate individual elements that make up a mixture signal and analyze them or use them or different functions ranging from re mixing, mastering or for educational purpose.With different instruments, sounds, timbers interacting with each other, it is difficult to visualize their combination to make the final mixture signal.There were few methods which attempted exploiting the statistical relations of the individual sources with final the final mixture signals.With the arrival of machine learning, neural networks, researchers are curious to know the outcome of applying various deep learning models for solving this problem of audio source separation. The availability of larger memory and processing power has encouraged the use of deep learning methodologies in solving various problems.Their ability find interesting patterns with the introduction of non linearity, convolutions layers, short memory cells has helped achieve better results in the domains of image, video, audio. These models are flexible, hence a model used in one domain can be modified to suite other domains as well. The development of various APIs like Tensorflow, Keras, Theano, Pytorch has made the realization and application of complicated operations involved in deep learning models easy to understand and implement. A song is made up of different sources, instruments. In this thesis our main focus would be to extract bass, drums and vocals from a given song.These three elemnts have distinct timber and also different frequency regions where they have maximum presence.These sources are also the driving force of a song. Different techniques have been used till date to solve this problem.An overview of these techniques, proposed model and the elements included are explained in the chapters ahead.